Contrastive Open Set Recognition

In conventional recognition tasks, models are only trained to recognize learned targets, but it is usually difficult collect training examples of all potential categories. the testing phase, when receive test samples from unknown classes, they mistakenly classify into known classes. Open set (OSR) a more realistic task, which requires classifier detect while keeping high classification accuracy this paper, we study how improve OSR performance deep neural networks perspective representation learning. We employ supervised contrastive learning quality feature representations, propose new method that enables model learn soft and design an framework on its basis. With proposed method, able make use label smoothing mixup contrastively, so as both robustness outlier detection in tasks tasks. validate our multiple benchmark datasets scenarios, achieving experimental results verify effectiveness method.